A streamed tool call without an index starts a new call when it brings an id

[gedaiu]

What happened

With #1183 in place, a question against Gemini (gemini-3.8-flash, OpenAI-compatible endpoint) still failed on its first turn. The log line from #1189 says why:

UnknownToolCall: model attempted to call unknown or disallowed tool `find_entitiessearch_chunks`

Gemini had made two parallel calls. Its stream sends them one per chunk, with no index on either and an id on each:

data: {"choices":[{"delta":{"tool_calls":[{"function":{"arguments":"{\"name\":\"lore\"}","name":"find_entities"},"id":"call_266956","type":"function"}]},"index":0}]}
data: {"choices":[{"delta":{"tool_calls":[{"function":{"arguments":"{\"query\":\"…\"}","name":"search_chunks"},"id":"call_266959","type":"function"}]},"index":0}]}

The streaming parser keyed each delta by index, reading a missing one as 0, so both calls merged into slot 0 and the names concatenated.

What changed

tool_call_slot decides where a delta goes: a numbered one where it says; an unnumbered one with an id starts a new call; an unnumbered one without an id continues the last call, which is the only shape an unnumbered continuation can have (OpenAI's continuation deltas carry argument fragments and no id). Numbered streams behave exactly as before.

Tests: the Gemini two-chunk shape yields two calls with their own ids, names and arguments; an unnumbered, id-less fragment appends to the previous call.

cargo fmt --check, cargo clippy --all-targets -- -D warnings, cargo test -p utopia-llm pass (49).

Related: #1135, #1183, #1189.

[WaylandYang]

Thank you for following the Gemini path to the end. The cause is clear from the two frames you quote, and the rule reads correctly: a numbered delta goes where it says, an unnumbered one with an id starts a call, an unnumbered one without continues the last. Checked here: 53 utopia-llm tests pass, and parallel_tool_calls_without_an_index_stay_separate fails without the change. One note, not blocking: the continuation test passes without the change too, since with a single call both rules pick slot 0; a second call ahead of the fragment would make it bite. Merging now. Thanks again.